Tag
Rollout-Decoded Reconstruction (RDR) is a loss term that improves long-horizon prediction in latent world models by free-running the model during training, achieving a 1.80× improvement in valid prediction time on chaotic systems.